Research Experience

Have the courage to follow your heart and intuition. ​——Steve Jobs

#Photographed at East Lake, Wuhan

Video Agent @ ByFusion (LuckyShort) — Agentic Spatio-Temporally Grounded Video Generation

At ByFusion (LuckyShort), where I serve as Co-founder and CTO, we develop agentic, spatio-temporally grounded systems for short-form episodic video generation. Multimodal agents autonomously script, storyboard, generate, and publish video end-to-end, coordinated by an orchestration layer that treats planning, asset generation, and rendering as cooperating processes. The central problems are controllable long-context generation and the spatio-temporal grounding needed to maintain narrative and visual consistency across episodes while limiting error accumulation. The resulting system now drives multiple fully automated video channels.

Agent-operated channels:  Video 1 · Channel 1 · Channel 2 · Channel 3 · Channel 4 · Channel 5 · Channel 6 · Channel 7 · Channel 8 · Channel 9 · Channel 10

Traffic-R1: Human-Like Reasoning for Traffic Signal Control [paper] (ACL 2026)

We present Traffic-R1, a lightweight 3B-parameter reasoning language model for traffic-signal control, trained with a two-stage reinforcement-learning framework. The model performs human-like, step-by-step reasoning over road observations, traffic incidents, and messages exchanged with neighboring intersections. It achieves zero-shot generalization to unseen road networks and out-of-distribution incidents, such as yielding to an ambulance, while remaining compact enough for edge deployment. This work appears in ACL 2026.

DeepUHI: Context-Aware Thermodynamic Modeling for Urban Heat-Island Forecasting [paper][code] (KDD 2025)

We propose DeepUHI, a context-aware thermodynamic modeling framework for fine-grained, street-level urban heat-island forecasting. The framework couples deep neural networks with thermodynamic domain knowledge to produce accurate and interpretable temperature predictions. We also introduce SeoulTemp, the first fine-grained urban-temperature dataset, comprising 947 stations across 605 km² of Seoul over the period 2021–2024. This work appears in KDD 2025.

GeoHG: Heterogeneous-Graph Learning of Geospatial Region Embeddings [paper][code]

We propose GeoHG, a heterogeneous-graph approach to learning geospatial region embeddings that integrates satellite-image features with points of interest and socio-environmental data. The method produces comprehensive region representations that generalize across regions, including settings with limited data availability. A preprint is available on arXiv.

Deep Learning for Cross-Domain Data Fusion in Urban Computing: Taxonomy, Advances, and Outlook  [paper] [ code]

We present the first systematic review of deep-learning-based cross-domain data fusion in urban computing. The survey proposes a taxonomy organized along data perspectives, methods, and applications, and discusses the emerging role of large language models in this setting. The work appears in Information Fusion (impact factor 14.9).

Model for Prediction of Rock Joint Roughness and Based Convolutional Neural Network [paper]

We develop a non-contact workflow that 3D-scans rock-joint surfaces and applies deep learning to quantify joint roughness (JRC) and predict shear strength, reaching JRC prediction accuracy above 75%. The method replaces labor-intensive physical measurement with a data-driven pipeline. Supervised by Dr. Qi Zhao, this study marked an early transition from physical experiments toward data-driven modeling. 

Smart Classifier for Worker's Working Posture for Rapid Entire Body Assessment (REBA)

We propose a CNN-LSTM time-series model that infers ergonomic-risk (REBA) scores from smartphone sensor data collected from construction workers. The approach enables scalable posture monitoring for occupational health without dedicated instrumentation. This work was supervised by  Dr. Yantao Yu.

Automated Solutions for Large-scale Waterproofing Construction; JBOT Limited (CR NO. 3223788 HK)

JBOT (HK) LIMITED (CR NO. 3223788). As co-founder, I led the development of a computer-vision-guided construction robot for large-scale waterproofing that reaches over 95% completion rate while saving materials. Its mechanical design and vision-alignment system, combined with pre-mixed combustion, are integrated into an IoT-enabled intelligent-construction system that links on-site perception and control with a broader monitoring infrastructure. 


Extreme Temperature Distribution in Concrete Bridges Under Climate Change [paper]
We modeled the extreme temperature distribution in large concrete bridges under a warming climate, combining Shenzhen Bay Bridge monitoring data and Hong Kong Observatory records with calibrated Abaqus finite-element thermodynamic models, informing structural-reinforcement recommendations for long-term bridge performance. The work received support from the Architectural Services Department of Hong Kong and formed my graduation thesis, under the guidance of  Prof. F.T.K. Au.


( To be updated  🙂)